QwenOFT-5traj (LIBERO-Goal)

Base Vision-Language-Action (VLA) model released with the AlphaBrain framework. Used as the frozen backbone for the RL-Token continual-learning experiments; also works standalone for LIBERO-Goal.

A QwenOFT VLA — Qwen2.5-VL-3B backbone with an MLP / Flow-Matching action head — fine-tuned on 5 trajectories per task across the 10 LIBERO-Goal tasks. Serves as the starting checkpoint for the companion RL-Token release AlphaBrainGroup/alphabrain-rlt-5traj-alltasks-libero-goal.

Overview


Architecture	QwenOFT (Qwen2.5-VL-3B + OFT action head)
Base VLM	`Qwen/Qwen2.5-VL-3B-Instruct`
Parameters	~3.8 B (all trainable during supervised fine-tune)
Training data	LIBERO-Goal, 10 tasks × 5 demonstrations each
Action head	Parallel continuous action decoding
Primary use	Frozen backbone for RL-Token fine-tuning (see companion repo)

Files

├── README.md                  model card
├── framework_config.yaml      AlphaBrain framework configuration
├── dataset_statistics.json    action normalisation statistics (required for inference)
├── model.safetensors          full VLA weights (~7.8 GB)
└── qwen_pretrained/           tokenizer, preprocessor, and Qwen2.5-VL config

Usage

git clone https://github.com/AlphaBrainGroup/AlphaBrain.git
cd VLA-Engine-Developer
pip install -e .

export PRETRAINED_MODELS_DIR=/path/to/models   # must contain Qwen2.5-VL-3B-Instruct/

huggingface-cli download AlphaBrainGroup/qwenoft-5traj-libero-goal \
    --local-dir ./qwenoft_5traj

# Launch the WebSocket inference server
python deployment/model_server/server_policy.py \
    --ckpt_path ./qwenoft_5traj --port 10093 --use_bf16

For LIBERO-Goal evaluation, see the framework's LIBERO eval instructions. To use this as the base for RL-Token fine-tuning, download both this repo and the RLT companion repo — see the RLT card for the full pipeline.

Companion release

AlphaBrainGroup/alphabrain-rlt-5traj-alltasks-libero-goal — RL-Token fine-tuned checkpoint that wraps this base VLA and achieves substantially higher success rates with tiny additional parameters.

License

MIT — see the parent repository.

Citation

@misc{alphabrain2026,
  title  = {AlphaBrain: A Modular Open-Source Framework for Embodied Intelligence Research},
  author = {AlphaBrain Team},
  year   = {2026},
  url    = {https://github.com/AlphaBrainGroup/AlphaBrain}
}